Mapping Crustal Vp/Vs in North America With a Machine Learning Approach

Abstract

Vp/Vs (Poisson's ratio) provides critical information for constraining the bulk crustal composition, stress state, and tectonic evolution of the Earth. The receiver function technique has been extensively utilized to constrain the crustal Vp/Vs, yet the reliability of measurements can be affected by complex structures and uneven distribution of seismic stations. Consequently, the interpolated Vp/Vs maps can often be biased by unreliable observations, especially in data-sparse regions. We tackle these issues by proposing a machine learning model that integrates multiple geophysical data sets to estimate Vp/Vs, leveraging the physical and structural properties of the crust. We train the model by compiling an extensive data set of global Vp/Vs measurements at 13,314 seismic stations and employ XGBoost to map Vp/Vs with other key crustal properties. Experiments using data from the (a) United States and (b) United States and Canada demonstrate superior prediction accuracy, achieving an overall $R^2$ value of 0.84 in both cases. Feature importance analysis indicates that crustal tectonic type, geographic coordinates, mid-crust shear-wave velocity, and crustal thickness primarily capture Vp/Vs variations, together explaining over 70% of reduction in the normalized root-mean-square error. The inclusion of other features further refines small-scale Vp/Vs variation. Compared to cubic and Kriging interpolations, the predicted Vp/Vs map from machine learning exhibits less local extremes and a better alignment with the first-order crustal structure across the continent. This study highlights the capability of machine learning to uncover complex geophysical relationships for reliable Vp/Vs estimates and its potential to constrain crustal composition at a continental scale.